Microsoft

Custom Vision Service

Welcome to the Custom Vision Service API Forum

Categories

API – Any ideas or feedback pertaining to features or enhancements to Custom Vision Service API.

Documentation – Any ideas or suggestions for the API Reference or Documentation.

Language Support – Submit a request to have a particular language supported.

Samples & SDK Request – Let us know if you would like to see a Code sample or SDK provided.

Release Notes – Recently updated December 19.

  • Hot ideas
  • Top ideas
  • New ideas
  • My feedback
  1. Train on images with no tags

    The docs suggest creating a fake tag for images with nothing in them. This is going to take space in the model and try for it to optimize something which doesn't make sense. When I get an image with a false positive I want to be able to upload the image to the training set even though there is nothing of interest in the picture.

    1 vote
    Sign in
    (thinking…)
    Sign in with: Facebook Google
    Signed in as (Sign out)

    We’ll send you updates on this idea

    0 comments  ·  Training  ·  Flag idea as inappropriate…  ·  Admin →
  2. Backup Project

    Ability to simply backup or even copy a project.
    The delete button is easily available, please add a copy button/feature.
    Thank you,
    Jared

    3 votes
    Sign in
    (thinking…)
    Sign in with: Facebook Google
    Signed in as (Sign out)

    We’ll send you updates on this idea

    0 comments  ·  Training  ·  Flag idea as inappropriate…  ·  Admin →
  3. Retail Domain for Object Detection

    We're working on a model to recognize clothe types for an app. We've tested both object detection and classification, and classification is about 6.5% more accurate, most likely because of the specialized Retail domain. However, this domain is not available for object detection.

    Can you add the pre-built retail domain to object detection projects?

    1 vote
    Sign in
    (thinking…)
    Sign in with: Facebook Google
    Signed in as (Sign out)

    We’ll send you updates on this idea

    0 comments  ·  Training  ·  Flag idea as inappropriate…  ·  Admin →
  4. Shortcuts Left Hand

    While training data, people use the right hand to draw boxes. Buttons near the left hand, tab, w-a-s-d need to be utilized for fast labelling. a-d for image right/left, w-s for up and down through labels.
    It makes labelling much much faster.

    1 vote
    Sign in
    (thinking…)
    Sign in with: Facebook Google
    Signed in as (Sign out)

    We’ll send you updates on this idea

    0 comments  ·  Training  ·  Flag idea as inappropriate…  ·  Admin →
  5. be able to zoom

    Please enable zooming in on an image to tag small features in training.

    5 votes
    Sign in
    (thinking…)
    Sign in with: Facebook Google
    Signed in as (Sign out)

    We’ll send you updates on this idea

    1 comment  ·  Training  ·  Flag idea as inappropriate…  ·  Admin →
  6. Allow retraining of image classification model via Python

    The custom vision documentation contains an example of how to create and train an image classification model then use it to make predictions, all in Python. It would be useful to be able to retrain already existing models using additional images in Python, as this can currently be done when viewing projects on the website but when attempting to retrain via Python just creates a new project with only the new images.
    There is also the issue that when training, training can only be completed when there are images present for all tags created, otherwise the error "CustomVisionErrorException: Not enough…

    1 vote
    Sign in
    (thinking…)
    Sign in with: Facebook Google
    Signed in as (Sign out)

    We’ll send you updates on this idea

    0 comments  ·  Training  ·  Flag idea as inappropriate…  ·  Admin →
  7. Select subset of data for Training

    I would like the ability to select a subset of images for Training, to enable my own cross-validation. Right now I don't think I have any control over which specific images are included in the model, unless I delete them (for later re-importing).

    2 votes
    Sign in
    (thinking…)
    Sign in with: Facebook Google
    Signed in as (Sign out)

    We’ll send you updates on this idea

    0 comments  ·  Training  ·  Flag idea as inappropriate…  ·  Admin →
  8. Ability to import iterations into the workspace to allow models to be generated on subsets of data

    I'd like to be able to Import the training data from an iteration into the workspace, modify that data and train a new iteration.

    This would allow multiple iterations to be trained with subsets of the same data. ie, you could have iterations containing data for Asia-Pacific, The Americas, and Global, all using a subset of the same training data. You can then import one of these iterations into your workspace to update the training images and create a new iteration for that locale, or import the global iteration to have access to the full set of training images.

    It…

    3 votes
    Sign in
    (thinking…)
    Sign in with: Facebook Google
    Signed in as (Sign out)

    We’ll send you updates on this idea

    Under Review  ·  0 comments  ·  Training  ·  Flag idea as inappropriate…  ·  Admin →
  9. Provide feature to download existing training images

    Provide feature to download existing training images

    4 votes
    Sign in
    (thinking…)
    Sign in with: Facebook Google
    Signed in as (Sign out)

    We’ll send you updates on this idea

    Under Review  ·  1 comment  ·  Training  ·  Flag idea as inappropriate…  ·  Admin →
  10. Train with Random cropping - Predict with multiple cropping and majority voting

    Have a feature to switch on cropping (data augmentation) for training and for inference.

    1) It is a well known technique to use random cropping to help to train a classifier. (Alex Krizhevsky did that too)

    2) In this paper https://arxiv.org/pdf/1803.00949v1.pdf about Tree Species Identification from Bark Images Using Convolutional
    Neural Networks researchers used cropping for training and for interference too. " For
    the latter, we split the test image into multiple non-overlapping
    224x224 crops and classified each one individually. Then, we
    performed majority voting to determine the final outcome.
    As can be seen from Table III, progressing from single …

    5 votes
    Sign in
    (thinking…)
    Sign in with: Facebook Google
    Signed in as (Sign out)

    We’ll send you updates on this idea

    Under Review  ·  0 comments  ·  Training  ·  Flag idea as inappropriate…  ·  Admin →
  11. 3 votes
    Sign in
    (thinking…)
    Sign in with: Facebook Google
    Signed in as (Sign out)

    We’ll send you updates on this idea

    Under Review  ·  2 comments  ·  Training  ·  Flag idea as inappropriate…  ·  Admin →
  12. Training data analysis for improvement

    After training, it would be nice if we could see analysis on biases of train data set (e.g. a set of data tagged to "A" seems similar than other tags, or seem to have enough deviation in dataset, etc), so tthat we could add new data to correct those.

    7 votes
    Sign in
    (thinking…)
    Sign in with: Facebook Google
    Signed in as (Sign out)

    We’ll send you updates on this idea

    Under Review  ·  1 comment  ·  Training  ·  Flag idea as inappropriate…  ·  Admin →
  13. Training - Allow us to crop or set bounds box for each tag

    I'd like the option to upload a single picture and tag it into multiple tags but tell the computer where in the picture to focus on what tag. This could help us get object level tagging if we wanted to take the time to train it.

    10 votes
    Sign in
    (thinking…)
    Sign in with: Facebook Google
    Signed in as (Sign out)

    We’ll send you updates on this idea

    Under Review  ·  4 comments  ·  Training  ·  Flag idea as inappropriate…  ·  Admin →
  14. Orientation of images

    The orientation of the image matter when using, I think it is better if there is an option to let the system know to change the orientation and train

    4 votes
    Sign in
    (thinking…)
    Sign in with: Facebook Google
    Signed in as (Sign out)

    We’ll send you updates on this idea

    Under Review  ·  0 comments  ·  Training  ·  Flag idea as inappropriate…  ·  Admin →
  15. Statistical tools

    It would be great to have access to more some more basic stats on the training process - learning curves, errors, ROCs, AUC values - the more the better.

    4 votes
    Sign in
    (thinking…)
    Sign in with: Facebook Google
    Signed in as (Sign out)

    We’ll send you updates on this idea

    Under Review  ·  1 comment  ·  Training  ·  Flag idea as inappropriate…  ·  Admin →
  16. Eliminate false positives

    I have trained my model with 500 images of different BMW motor vehicles however if i give it an image of a stretch Hummer it categorizes that as a BMW.

    It is able to correctly return low probability (or zero) for clearly non car related imagery but i would have thought it should also return a low probability of a stretch Hummer being a BMW too, not 99%

    Put another way, if an image is of the same type as that which the model is trained for then unless you have sufficient tags (e.g, lots of different vehicle makes) then…

    8 votes
    Sign in
    (thinking…)
    Sign in with: Facebook Google
    Signed in as (Sign out)

    We’ll send you updates on this idea

    Under Review  ·  3 comments  ·  Training  ·  Flag idea as inappropriate…  ·  Admin →
  17. Ignore tags with less than 5 images when training

    Currently, if you want to train your project, all tags need to have at least 5 images otherwise an error will be thrown and your entire project will not get trained.

    Please allow an option to ignore tags with less that 5 images when training

    11 votes
    Sign in
    (thinking…)
    Sign in with: Facebook Google
    Signed in as (Sign out)

    We’ll send you updates on this idea

    Under Review  ·  5 comments  ·  Training  ·  Flag idea as inappropriate…  ·  Admin →
  18. Add the ability to train the api / provide feedback

    I have an application that uses the API to do OCR, but I want to be able to provide feedback every time it is used so that the 'correct' answer can be fed back into the AI engine so that the engine can be improved over time.

    I have a second application that will in the future use tags, and I could also provide feedback to the engine as to what the user actually selected after the suggested tags from the API were provided.

    10 votes
    Sign in
    (thinking…)
    Sign in with: Facebook Google
    Signed in as (Sign out)

    We’ll send you updates on this idea

    Resolved  ·  2 comments  ·  Training  ·  Flag idea as inappropriate…  ·  Admin →
  19. Train the vision api with custom tags/domain specific models

    I would like the ability to train the API with custom tags.

    For instance, I would like to feed in pictures of cat breeds to 'train' the AI, then receive the breed value on an AI request.

    49 votes
    Sign in
    (thinking…)
    Sign in with: Facebook Google
    Signed in as (Sign out)

    We’ll send you updates on this idea

    11 comments  ·  Training  ·  Flag idea as inappropriate…  ·  Admin →
    Completed  ·  Raymond responded

    We now have this type of capability using our Custom Vision Service. You can find further details to get started on the Custom Vision Service overview page.

    https://docs.microsoft.com/en-us/azure/cognitive-services/custom-vision-service/home

    Custom Vision Service is a tool for building custom image classifiers, and for making them better over time. For example, if you want a tool that could identify images of “Daisies”, “Daffodils”, and “Dahlias”, you could train a classifier to do that. You do so by providing Custom Vision Service with images for each tag you want to recognize.

  • Don't see your idea?

Feedback and Knowledge Base